Most current approaches to understanding the neural basis of cognitive processes are severely limited in two respects. First, most commonly used methods do not have the temporal (e.g., fMRI) or spatial (e.g., MEG/ EEG) resolution to capture the relevant dynamics. Second, even methods with high spatio-temporal resolution (intracranial EEG - icEEG) typically approach target cognitive processes in a fragmentary, un- integrated way. For instance, language is typically studied as a conglomeration of separate subsystems: perception, pattern recognition, categorization, semantically/syntactically appropriate response selection, cross-modal integration, motor control and sensorimotor integration. The present proposal aims to remedy both limitations by using icEEG to study a model system, reading/speech/language, from an integrative and unified perspective. We focus on reading, a complex task that involves visual pattern recognition, visual- auditory and visuo-motor integration, semantic, syntactic and phonological access, and (in reading aloud) - response selection and motor sequencing. Reading allows for easy, yet ecologically valid manipulations of cognitive load in the language system. The neuro-computational framework we propose to test is that computation is achieved not by information passing through a sequence of discrete processing stages in individual modules but via state transitions of a distributed network. We will recruit a large cohort of 80 patients in whom we will quantify both local as well as inter-regional cortical dynamics during word reading - from early primary visual perception, through selection, to word output. We will leverage our established techniques for precise co-localization and analysis of grouped icEEG data, circumventing the sparse sampling problem inherent to human icEEG experiments. The combined use of sub-dural grid electrodes and stereo-electroencephalographic depth electrodes will enable the study of not only classic peri-sylvian regions, but also of deep sulci (and regions such as the planum temporale). We will then characterize dynamic network interactions using linear and non linear measures of amplitude covariance in high frequencies, following analyses we have developed previously. Critical nodes and critical transitions in network states will then be perturbed using closed-loop activity-triggered direct cortical stimulation. To achieve these goals we have set up a collaboration between the Texas Comprehensive Epilepsy Program and Johns Hopkins Medical Center - both centers have a proven record of studying language with icEEG. Our team has expertise in all aspects language, reading, icEEG signal analysis, population level network modeling from intracranial recordings; and neural networks. This work will dramatically improve our understanding of language systems and test and develop a new way to model neural computation generally.